Refactor Endpoint API

- Allow reads to complete immediately
- Unify read/write signatures
- Simplify memory management to allow future optimization work

3 files changed, 206 insertions, 382 deletions

diff --git a/src/core/iomgr/endpoint.c b/src/core/iomgr/endpoint.c
index 8ee14bce9b..a7878e31dd 100644
--- a/src/core/iomgr/endpoint.c
+++ b/src/core/iomgr/endpoint.c
@@ -33,17 +33,16 @@
 
 #include "src/core/iomgr/endpoint.h"
 
-void grpc_endpoint_notify_on_read(grpc_endpoint *ep, grpc_endpoint_read_cb cb,
-                                  void *user_data) {
-  ep->vtable->notify_on_read(ep, cb, user_data);
+grpc_endpoint_op_status grpc_endpoint_read(grpc_endpoint *ep,
+                                           gpr_slice_buffer *slices,
+                                           grpc_iomgr_closure *cb) {
+  return ep->vtable->read(ep, slices, cb);
 }
 
-grpc_endpoint_write_status grpc_endpoint_write(grpc_endpoint *ep,
-                                               gpr_slice *slices,
-                                               size_t nslices,
-                                               grpc_endpoint_write_cb cb,
-                                               void *user_data) {
-  return ep->vtable->write(ep, slices, nslices, cb, user_data);
+grpc_endpoint_op_status grpc_endpoint_write(grpc_endpoint *ep,
+                                            gpr_slice_buffer *slices,
+                                            grpc_iomgr_closure *cb) {
+  return ep->vtable->write(ep, slices, cb);
 }
 
 void grpc_endpoint_add_to_pollset(grpc_endpoint *ep, grpc_pollset *pollset) {
diff --git a/src/core/iomgr/endpoint.h b/src/core/iomgr/endpoint.h
index ea92a500e8..38f1e46d67 100644
--- a/src/core/iomgr/endpoint.h
+++ b/src/core/iomgr/endpoint.h
@@ -37,6 +37,7 @@
 #include "src/core/iomgr/pollset.h"
 #include "src/core/iomgr/pollset_set.h"
 #include <grpc/support/slice.h>
+#include <grpc/support/slice_buffer.h>
 #include <grpc/support/time.h>
 
 /* An endpoint caps a streaming channel between two communicating processes.
@@ -45,31 +46,17 @@
 typedef struct grpc_endpoint grpc_endpoint;
 typedef struct grpc_endpoint_vtable grpc_endpoint_vtable;
 
-typedef enum grpc_endpoint_cb_status {
-  GRPC_ENDPOINT_CB_OK = 0, /* Call completed successfully */
-  GRPC_ENDPOINT_CB_EOF, /* Call completed successfully, end of file reached */
-  GRPC_ENDPOINT_CB_SHUTDOWN, /* Call interrupted by shutdown */
-  GRPC_ENDPOINT_CB_ERROR     /* Call interrupted by socket error */
-} grpc_endpoint_cb_status;
-
-typedef enum grpc_endpoint_write_status {
-  GRPC_ENDPOINT_WRITE_DONE,    /* completed immediately, cb won't be called */
-  GRPC_ENDPOINT_WRITE_PENDING, /* cb will be called when completed */
-  GRPC_ENDPOINT_WRITE_ERROR    /* write errored out, cb won't be called */
-} grpc_endpoint_write_status;
-
-typedef void (*grpc_endpoint_read_cb)(void *user_data, gpr_slice *slices,
-                                      size_t nslices,
-                                      grpc_endpoint_cb_status error);
-typedef void (*grpc_endpoint_write_cb)(void *user_data,
-                                       grpc_endpoint_cb_status error);
+typedef enum grpc_endpoint_op_status {
+  GRPC_ENDPOINT_DONE,    /* completed immediately, cb won't be called */
+  GRPC_ENDPOINT_PENDING, /* cb will be called when completed */
+  GRPC_ENDPOINT_ERROR    /* write errored out, cb won't be called */
+} grpc_endpoint_op_status;
 
 struct grpc_endpoint_vtable {
-  void (*notify_on_read)(grpc_endpoint *ep, grpc_endpoint_read_cb cb,
-                         void *user_data);
-  grpc_endpoint_write_status (*write)(grpc_endpoint *ep, gpr_slice *slices,
-                                      size_t nslices, grpc_endpoint_write_cb cb,
-                                      void *user_data);
+  grpc_endpoint_op_status (*read)(grpc_endpoint *ep, gpr_slice_buffer *slices,
+                                  grpc_iomgr_closure *cb);
+  grpc_endpoint_op_status (*write)(grpc_endpoint *ep, gpr_slice_buffer *slices,
+                                   grpc_iomgr_closure *cb);
   void (*add_to_pollset)(grpc_endpoint *ep, grpc_pollset *pollset);
   void (*add_to_pollset_set)(grpc_endpoint *ep, grpc_pollset_set *pollset);
   void (*shutdown)(grpc_endpoint *ep);
@@ -77,9 +64,13 @@ struct grpc_endpoint_vtable {
   char *(*get_peer)(grpc_endpoint *ep);
 };
 
-/* When data is available on the connection, calls the callback with slices. */
-void grpc_endpoint_notify_on_read(grpc_endpoint *ep, grpc_endpoint_read_cb cb,
-                                  void *user_data);
+/* When data is available on the connection, calls the callback with slices.
+   Callback success indicates that the endpoint can accept more reads, failure
+   indicates the endpoint is closed.
+   Valid slices may be placed into \a slices even on callback success == 0. */
+grpc_endpoint_op_status grpc_endpoint_read(
+    grpc_endpoint *ep, gpr_slice_buffer *slices,
+    grpc_iomgr_closure *cb) GRPC_MUST_USE_RESULT;
 
 char *grpc_endpoint_get_peer(grpc_endpoint *ep);
 
@@ -89,11 +80,9 @@ char *grpc_endpoint_get_peer(grpc_endpoint *ep);
    returns GRPC_ENDPOINT_WRITE_DONE.
    Otherwise it returns GRPC_ENDPOINT_WRITE_PENDING and calls cb when the
    connection is ready for more data. */
-grpc_endpoint_write_status grpc_endpoint_write(grpc_endpoint *ep,
-                                               gpr_slice *slices,
-                                               size_t nslices,
-                                               grpc_endpoint_write_cb cb,
-                                               void *user_data);
+grpc_endpoint_op_status grpc_endpoint_write(
+    grpc_endpoint *ep, gpr_slice_buffer *slices,
+    grpc_iomgr_closure *cb) GRPC_MUST_USE_RESULT;
 
 /* Causes any pending read/write callbacks to run immediately with
    GRPC_ENDPOINT_CB_SHUTDOWN status */
diff --git a/src/core/iomgr/tcp_posix.c b/src/core/iomgr/tcp_posix.c
index 360e6ebd8c..36ba3a7606 100644
--- a/src/core/iomgr/tcp_posix.c
+++ b/src/core/iomgr/tcp_posix.c
@@ -61,209 +61,8 @@
 #define SENDMSG_FLAGS 0
 #endif
 
-/* Holds a slice array and associated state. */
-typedef struct grpc_tcp_slice_state {
-  gpr_slice *slices;       /* Array of slices */
-  size_t nslices;          /* Size of slices array. */
-  ssize_t first_slice;     /* First valid slice in array */
-  ssize_t last_slice;      /* Last valid slice in array */
-  gpr_slice working_slice; /* pointer to original final slice */
-  int working_slice_valid; /* True if there is a working slice */
-  int memory_owned;        /* True if slices array is owned */
-} grpc_tcp_slice_state;
-
 int grpc_tcp_trace = 0;
 
-static void slice_state_init(grpc_tcp_slice_state *state, gpr_slice *slices,
-                             size_t nslices, size_t valid_slices) {
-  state->slices = slices;
-  state->nslices = nslices;
-  if (valid_slices == 0) {
-    state->first_slice = -1;
-  } else {
-    state->first_slice = 0;
-  }
-  state->last_slice = valid_slices - 1;
-  state->working_slice_valid = 0;
-  state->memory_owned = 0;
-}
-
-/* Returns true if there is still available data */
-static int slice_state_has_available(grpc_tcp_slice_state *state) {
-  return state->first_slice != -1 && state->last_slice >= state->first_slice;
-}
-
-static ssize_t slice_state_slices_allocated(grpc_tcp_slice_state *state) {
-  if (state->first_slice == -1) {
-    return 0;
-  } else {
-    return state->last_slice - state->first_slice + 1;
-  }
-}
-
-static void slice_state_realloc(grpc_tcp_slice_state *state, size_t new_size) {
-  /* TODO(klempner): use realloc instead when first_slice is 0 */
-  /* TODO(klempner): Avoid a realloc in cases where it is unnecessary */
-  gpr_slice *slices = state->slices;
-  size_t original_size = slice_state_slices_allocated(state);
-  size_t i;
-  gpr_slice *new_slices = gpr_malloc(sizeof(gpr_slice) * new_size);
-
-  for (i = 0; i < original_size; ++i) {
-    new_slices[i] = slices[i + state->first_slice];
-  }
-
-  state->slices = new_slices;
-  state->last_slice = original_size - 1;
-  if (original_size > 0) {
-    state->first_slice = 0;
-  } else {
-    state->first_slice = -1;
-  }
-  state->nslices = new_size;
-
-  if (state->memory_owned) {
-    gpr_free(slices);
-  }
-  state->memory_owned = 1;
-}
-
-static void slice_state_remove_prefix(grpc_tcp_slice_state *state,
-                                      size_t prefix_bytes) {
-  gpr_slice *current_slice = &state->slices[state->first_slice];
-  size_t current_slice_size;
-
-  while (slice_state_has_available(state)) {
-    current_slice_size = GPR_SLICE_LENGTH(*current_slice);
-    if (current_slice_size > prefix_bytes) {
-      /* TODO(klempner): Get rid of the extra refcount created here by adding a
-         native "trim the first N bytes" operation to splice */
-      /* TODO(klempner): This really shouldn't be modifying the current slice
-         unless we own the slices array. */
-      gpr_slice tail;
-      tail = gpr_slice_split_tail(current_slice, prefix_bytes);
-      gpr_slice_unref(*current_slice);
-      *current_slice = tail;
-      return;
-    } else {
-      gpr_slice_unref(*current_slice);
-      ++state->first_slice;
-      ++current_slice;
-      prefix_bytes -= current_slice_size;
-    }
-  }
-}
-
-static void slice_state_destroy(grpc_tcp_slice_state *state) {
-  while (slice_state_has_available(state)) {
-    gpr_slice_unref(state->slices[state->first_slice]);
-    ++state->first_slice;
-  }
-
-  if (state->memory_owned) {
-    gpr_free(state->slices);
-    state->memory_owned = 0;
-  }
-}
-
-void slice_state_transfer_ownership(grpc_tcp_slice_state *state,
-                                    gpr_slice **slices, size_t *nslices) {
-  *slices = state->slices + state->first_slice;
-  *nslices = state->last_slice - state->first_slice + 1;
-
-  state->first_slice = -1;
-  state->last_slice = -1;
-}
-
-/* Fills iov with the first min(iov_size, available) slices, returns number
-   filled */
-static size_t slice_state_to_iovec(grpc_tcp_slice_state *state,
-                                   struct iovec *iov, size_t iov_size) {
-  size_t nslices = state->last_slice - state->first_slice + 1;
-  gpr_slice *slices = state->slices + state->first_slice;
-  size_t i;
-  if (nslices < iov_size) {
-    iov_size = nslices;
-  }
-
-  for (i = 0; i < iov_size; ++i) {
-    iov[i].iov_base = GPR_SLICE_START_PTR(slices[i]);
-    iov[i].iov_len = GPR_SLICE_LENGTH(slices[i]);
-  }
-  return iov_size;
-}
-
-/* Makes n blocks available at the end of state, writes them into iov, and
-   returns the number of bytes allocated */
-static size_t slice_state_append_blocks_into_iovec(grpc_tcp_slice_state *state,
-                                                   struct iovec *iov, size_t n,
-                                                   size_t slice_size) {
-  size_t target_size;
-  size_t i;
-  size_t allocated_bytes;
-  ssize_t allocated_slices = slice_state_slices_allocated(state);
-
-  if (n - state->working_slice_valid >= state->nslices - state->last_slice) {
-    /* Need to grow the slice array */
-    target_size = state->nslices;
-    do {
-      target_size = target_size * 2;
-    } while (target_size < allocated_slices + n - state->working_slice_valid);
-    /* TODO(klempner): If this ever needs to support both prefix removal and
-       append, we should be smarter about the growth logic here */
-    slice_state_realloc(state, target_size);
-  }
-
-  i = 0;
-  allocated_bytes = 0;
-
-  if (state->working_slice_valid) {
-    iov[0].iov_base = GPR_SLICE_END_PTR(state->slices[state->last_slice]);
-    iov[0].iov_len = GPR_SLICE_LENGTH(state->working_slice) -
-                     GPR_SLICE_LENGTH(state->slices[state->last_slice]);
-    allocated_bytes += iov[0].iov_len;
-    ++i;
-    state->slices[state->last_slice] = state->working_slice;
-    state->working_slice_valid = 0;
-  }
-
-  for (; i < n; ++i) {
-    ++state->last_slice;
-    state->slices[state->last_slice] = gpr_slice_malloc(slice_size);
-    iov[i].iov_base = GPR_SLICE_START_PTR(state->slices[state->last_slice]);
-    iov[i].iov_len = slice_size;
-    allocated_bytes += slice_size;
-  }
-  if (state->first_slice == -1) {
-    state->first_slice = 0;
-  }
-  return allocated_bytes;
-}
-
-/* Remove the last n bytes from state */
-/* TODO(klempner): Consider having this defer actual deletion until later */
-static void slice_state_remove_last(grpc_tcp_slice_state *state, size_t bytes) {
-  while (bytes > 0 && slice_state_has_available(state)) {
-    if (GPR_SLICE_LENGTH(state->slices[state->last_slice]) > bytes) {
-      state->working_slice = state->slices[state->last_slice];
-      state->working_slice_valid = 1;
-      /* TODO(klempner): Combine these into a single operation that doesn't need
-         to refcount */
-      gpr_slice_unref(gpr_slice_split_tail(
-          &state->slices[state->last_slice],
-          GPR_SLICE_LENGTH(state->slices[state->last_slice]) - bytes));
-      bytes = 0;
-    } else {
-      bytes -= GPR_SLICE_LENGTH(state->slices[state->last_slice]);
-      gpr_slice_unref(state->slices[state->last_slice]);
-      --state->last_slice;
-      if (state->last_slice == -1) {
-        state->first_slice = -1;
-      }
-    }
-  }
-}
-
 typedef struct {
   grpc_endpoint base;
   grpc_fd *em_fd;
@@ -273,12 +72,15 @@ typedef struct {
   size_t slice_size;
   gpr_refcount refcount;
 
-  grpc_endpoint_read_cb read_cb;
-  void *read_user_data;
-  grpc_endpoint_write_cb write_cb;
-  void *write_user_data;
+  gpr_slice_buffer *incoming_buffer;
+  gpr_slice_buffer *outgoing_buffer;
+  /** slice within outgoing_buffer to write next */
+  size_t outgoing_slice_idx;
+  /** byte within outgoing_buffer->slices[outgoing_slice_idx] to write next */
+  size_t outgoing_byte_idx;
 
-  grpc_tcp_slice_state write_state;
+  grpc_iomgr_closure *read_cb;
+  grpc_iomgr_closure *write_cb;
 
   grpc_iomgr_closure read_closure;
   grpc_iomgr_closure write_closure;
@@ -288,65 +90,95 @@ typedef struct {
   char *peer_string;
 } grpc_tcp;
 
-static void grpc_tcp_handle_read(void *arg /* grpc_tcp */, int success);
-static void grpc_tcp_handle_write(void *arg /* grpc_tcp */, int success);
+static void tcp_handle_read(void *arg /* grpc_tcp */, int success);
+static void tcp_handle_write(void *arg /* grpc_tcp */, int success);
 
-static void grpc_tcp_shutdown(grpc_endpoint *ep) {
+static void tcp_shutdown(grpc_endpoint *ep) {
   grpc_tcp *tcp = (grpc_tcp *)ep;
   grpc_fd_shutdown(tcp->em_fd);
 }
 
-static void grpc_tcp_unref(grpc_tcp *tcp) {
-  int refcount_zero = gpr_unref(&tcp->refcount);
-  if (refcount_zero) {
-    grpc_fd_orphan(tcp->em_fd, NULL, "tcp_unref_orphan");
-    gpr_free(tcp->peer_string);
-    gpr_free(tcp);
+static void tcp_free(grpc_tcp *tcp) {
+  grpc_fd_orphan(tcp->em_fd, NULL, "tcp_unref_orphan");
+  gpr_free(tcp->peer_string);
+  gpr_free(tcp);
+}
+
+#define GRPC_TCP_REFCOUNT_DEBUG
+#define TCP_UNREF(tcp, reason) tcp_unref((tcp), (reason), __FILE__, __LINE__)
+#define TCP_REF(tcp, reason) tcp_ref((tcp), (reason), __FILE__, __LINE__)
+static void tcp_unref(grpc_tcp *tcp, const char *reason, const char *file,
+                      int line) {
+  gpr_log(file, line, GPR_LOG_SEVERITY_DEBUG, "TCP unref %p : %s %d -> %d", tcp,
+          reason, tcp->refcount.count, tcp->refcount.count - 1);
+  if (gpr_unref(&tcp->refcount)) {
+    tcp_free(tcp);
+  }
+}
+
+static void tcp_ref(grpc_tcp *tcp, const char *reason, const char *file,
+                    int line) {
+  gpr_log(file, line, GPR_LOG_SEVERITY_DEBUG, "TCP   ref %p : %s %d -> %d", tcp,
+          reason, tcp->refcount.count, tcp->refcount.count + 1);
+  gpr_ref(&tcp->refcount);
+}
+#ifdef GRPC_TCP_REFCOUNT_DEBUG
+#else
+#define TCP_UNREF(tcp, reason) tcp_unref((tcp))
+#define TCP_REF(tcp, reason) tcp_ref((tcp))
+static void tcp_unref(grpc_tcp *tcp) {
+  if (gpr_unref(&tcp->refcount)) {
+    tcp_free(tcp);
   }
 }
 
-static void grpc_tcp_destroy(grpc_endpoint *ep) {
+static void tcp_ref(grpc_tcp *tcp) { gpr_ref(&tcp->refcount); }
+#endif
+
+static void tcp_destroy(grpc_endpoint *ep) {
   grpc_tcp *tcp = (grpc_tcp *)ep;
-  grpc_tcp_unref(tcp);
+  TCP_UNREF(tcp, "destroy");
 }
 
-static void call_read_cb(grpc_tcp *tcp, gpr_slice *slices, size_t nslices,
-                         grpc_endpoint_cb_status status) {
-  grpc_endpoint_read_cb cb = tcp->read_cb;
+static void call_read_cb(grpc_tcp *tcp, int success) {
+  grpc_iomgr_closure *cb = tcp->read_cb;
 
   if (grpc_tcp_trace) {
     size_t i;
-    gpr_log(GPR_DEBUG, "read: status=%d", status);
-    for (i = 0; i < nslices; i++) {
-      char *dump = gpr_dump_slice(slices[i], GPR_DUMP_HEX | GPR_DUMP_ASCII);
+    gpr_log(GPR_DEBUG, "read: success=%d", success);
+    for (i = 0; i < tcp->incoming_buffer->count; i++) {
+      char *dump = gpr_dump_slice(tcp->incoming_buffer->slices[i],
+                                  GPR_DUMP_HEX | GPR_DUMP_ASCII);
       gpr_log(GPR_DEBUG, "READ %p: %s", tcp, dump);
       gpr_free(dump);
     }
   }
 
   tcp->read_cb = NULL;
-  cb(tcp->read_user_data, slices, nslices, status);
+  tcp->incoming_buffer = NULL;
+  cb->cb(cb->cb_arg, success);
 }
 
-#define INLINE_SLICE_BUFFER_SIZE 8
 #define MAX_READ_IOVEC 4
-static void grpc_tcp_continue_read(grpc_tcp *tcp) {
-  gpr_slice static_read_slices[INLINE_SLICE_BUFFER_SIZE];
+static void tcp_continue_read(grpc_tcp *tcp) {
   struct msghdr msg;
   struct iovec iov[MAX_READ_IOVEC];
   ssize_t read_bytes;
-  ssize_t allocated_bytes;
-  struct grpc_tcp_slice_state read_state;
-  gpr_slice *final_slices;
-  size_t final_nslices;
+  size_t i;
 
   GPR_ASSERT(!tcp->finished_edge);
+  GPR_ASSERT(tcp->iov_size <= MAX_READ_IOVEC);
+  GPR_ASSERT(tcp->incoming_buffer->count <= MAX_READ_IOVEC);
   GRPC_TIMER_BEGIN(GRPC_PTAG_HANDLE_READ, 0);
-  slice_state_init(&read_state, static_read_slices, INLINE_SLICE_BUFFER_SIZE,
-                   0);
 
-  allocated_bytes = slice_state_append_blocks_into_iovec(
-      &read_state, iov, tcp->iov_size, tcp->slice_size);
+  while (tcp->incoming_buffer->count < (size_t)tcp->iov_size) {
+    gpr_slice_buffer_add_indexed(tcp->incoming_buffer,
+                                 gpr_slice_malloc(tcp->slice_size));
+  }
+  for (i = 0; i < tcp->incoming_buffer->count; i++) {
+    iov[i].iov_base = GPR_SLICE_START_PTR(tcp->incoming_buffer->slices[i]);
+    iov[i].iov_len = GPR_SLICE_LENGTH(tcp->incoming_buffer->slices[i]);
+  }
 
   msg.msg_name = NULL;
   msg.msg_namelen = 0;
@@ -362,87 +194,63 @@ static void grpc_tcp_continue_read(grpc_tcp *tcp) {
   } while (read_bytes < 0 && errno == EINTR);
   GRPC_TIMER_END(GRPC_PTAG_RECVMSG, 0);
 
-  if (read_bytes < allocated_bytes) {
-    /* TODO(klempner): Consider a second read first, in hopes of getting a
-     * quick EAGAIN and saving a bunch of allocations. */
-    slice_state_remove_last(&read_state, read_bytes < 0
-                                             ? allocated_bytes
-                                             : allocated_bytes - read_bytes);
-  }
-
   if (read_bytes < 0) {
-    /* NB: After calling the user_cb a parallel call of the read handler may
+    /* NB: After calling call_read_cb a parallel call of the read handler may
      * be running. */
     if (errno == EAGAIN) {
       if (tcp->iov_size > 1) {
         tcp->iov_size /= 2;
       }
-      if (slice_state_has_available(&read_state)) {
-        /* TODO(klempner): We should probably do the call into the application
-           without all this junk on the stack */
-        /* FIXME(klempner): Refcount properly */
-        slice_state_transfer_ownership(&read_state, &final_slices,
-                                       &final_nslices);
-        tcp->finished_edge = 1;
-        call_read_cb(tcp, final_slices, final_nslices, GRPC_ENDPOINT_CB_OK);
-        slice_state_destroy(&read_state);
-        grpc_tcp_unref(tcp);
-      } else {
-        /* We've consumed the edge, request a new one */
-        slice_state_destroy(&read_state);
-        grpc_fd_notify_on_read(tcp->em_fd, &tcp->read_closure);
-      }
+      /* We've consumed the edge, request a new one */
+      grpc_fd_notify_on_read(tcp->em_fd, &tcp->read_closure);
     } else {
       /* TODO(klempner): Log interesting errors */
-      call_read_cb(tcp, NULL, 0, GRPC_ENDPOINT_CB_ERROR);
-      slice_state_destroy(&read_state);
-      grpc_tcp_unref(tcp);
+      gpr_slice_buffer_reset_and_unref(tcp->incoming_buffer);
+      call_read_cb(tcp, 0);
+      TCP_UNREF(tcp, "read");
     }
   } else if (read_bytes == 0) {
     /* 0 read size ==> end of stream */
-    if (slice_state_has_available(&read_state)) {
-      /* there were bytes already read: pass them up to the application */
-      slice_state_transfer_ownership(&read_state, &final_slices,
-                                     &final_nslices);
-      call_read_cb(tcp, final_slices, final_nslices, GRPC_ENDPOINT_CB_EOF);
-    } else {
-      call_read_cb(tcp, NULL, 0, GRPC_ENDPOINT_CB_EOF);
-    }
-    slice_state_destroy(&read_state);
-    grpc_tcp_unref(tcp);
+    gpr_slice_buffer_reset_and_unref(tcp->incoming_buffer);
+    call_read_cb(tcp, 0);
+    TCP_UNREF(tcp, "read");
   } else {
-    if (tcp->iov_size < MAX_READ_IOVEC) {
+    GPR_ASSERT((size_t)read_bytes <= tcp->incoming_buffer->length);
+    if ((size_t)read_bytes < tcp->incoming_buffer->length) {
+      gpr_slice_buffer_trim_end(tcp->incoming_buffer,
+                                tcp->incoming_buffer->length - read_bytes);
+    } else if (tcp->iov_size < MAX_READ_IOVEC) {
       ++tcp->iov_size;
     }
-    GPR_ASSERT(slice_state_has_available(&read_state));
-    slice_state_transfer_ownership(&read_state, &final_slices, &final_nslices);
-    call_read_cb(tcp, final_slices, final_nslices, GRPC_ENDPOINT_CB_OK);
-    slice_state_destroy(&read_state);
-    grpc_tcp_unref(tcp);
+    GPR_ASSERT((size_t)read_bytes == tcp->incoming_buffer->length);
+    call_read_cb(tcp, 1);
+    TCP_UNREF(tcp, "read");
   }
 
   GRPC_TIMER_END(GRPC_PTAG_HANDLE_READ, 0);
 }
 
-static void grpc_tcp_handle_read(void *arg /* grpc_tcp */, int success) {
+static void tcp_handle_read(void *arg /* grpc_tcp */, int success) {
   grpc_tcp *tcp = (grpc_tcp *)arg;
   GPR_ASSERT(!tcp->finished_edge);
 
   if (!success) {
-    call_read_cb(tcp, NULL, 0, GRPC_ENDPOINT_CB_SHUTDOWN);
-    grpc_tcp_unref(tcp);
+    call_read_cb(tcp, 0);
+    TCP_UNREF(tcp, "read");
   } else {
-    grpc_tcp_continue_read(tcp);
+    tcp_continue_read(tcp);
   }
 }
 
-static void grpc_tcp_notify_on_read(grpc_endpoint *ep, grpc_endpoint_read_cb cb,
-                                    void *user_data) {
+static grpc_endpoint_op_status tcp_read(grpc_endpoint *ep,
+                                        gpr_slice_buffer *incoming_buffer,
+                                        grpc_iomgr_closure *cb) {
   grpc_tcp *tcp = (grpc_tcp *)ep;
   GPR_ASSERT(tcp->read_cb == NULL);
   tcp->read_cb = cb;
-  tcp->read_user_data = user_data;
-  gpr_ref(&tcp->refcount);
+  tcp->incoming_buffer = incoming_buffer;
+  gpr_slice_buffer_reset_and_unref(incoming_buffer);
+  TCP_REF(tcp, "read");
   if (tcp->finished_edge) {
     tcp->finished_edge = 0;
     grpc_fd_notify_on_read(tcp->em_fd, &tcp->read_closure);
@@ -450,18 +258,41 @@ static void grpc_tcp_notify_on_read(grpc_endpoint *ep, grpc_endpoint_read_cb cb,
     tcp->handle_read_closure.cb_arg = tcp;
     grpc_iomgr_add_delayed_callback(&tcp->handle_read_closure, 1);
   }
+  /* TODO(ctiller): immediate return */
+  return GRPC_ENDPOINT_PENDING;
 }
 
 #define MAX_WRITE_IOVEC 16
-static grpc_endpoint_write_status grpc_tcp_flush(grpc_tcp *tcp) {
+static grpc_endpoint_op_status tcp_flush(grpc_tcp *tcp) {
   struct msghdr msg;
   struct iovec iov[MAX_WRITE_IOVEC];
   int iov_size;
   ssize_t sent_length;
-  grpc_tcp_slice_state *state = &tcp->write_state;
+  ssize_t sending_length;
+  ssize_t trailing;
+  ssize_t unwind_slice_idx;
+  ssize_t unwind_byte_idx;
 
   for (;;) {
-    iov_size = slice_state_to_iovec(state, iov, MAX_WRITE_IOVEC);
+    sending_length = 0;
+    unwind_slice_idx = tcp->outgoing_slice_idx;
+    unwind_byte_idx = tcp->outgoing_byte_idx;
+    for (iov_size = 0; tcp->outgoing_slice_idx != tcp->outgoing_buffer->count &&
+                       iov_size != MAX_WRITE_IOVEC;
+         iov_size++) {
+      iov[iov_size].iov_base =
+          GPR_SLICE_START_PTR(
+              tcp->outgoing_buffer->slices[tcp->outgoing_slice_idx]) +
+          tcp->outgoing_byte_idx;
+      iov[iov_size].iov_len =
+          GPR_SLICE_LENGTH(
+              tcp->outgoing_buffer->slices[tcp->outgoing_slice_idx]) -
+          tcp->outgoing_byte_idx;
+      sending_length += iov[iov_size].iov_len;
+      tcp->outgoing_slice_idx++;
+      tcp->outgoing_byte_idx = 0;
+    }
+    GPR_ASSERT(iov_size > 0);
 
     msg.msg_name = NULL;
     msg.msg_namelen = 0;
@@ -480,70 +311,75 @@ static grpc_endpoint_write_status grpc_tcp_flush(grpc_tcp *tcp) {
 
     if (sent_length < 0) {
       if (errno == EAGAIN) {
-        return GRPC_ENDPOINT_WRITE_PENDING;
+        tcp->outgoing_slice_idx = unwind_slice_idx;
+        tcp->outgoing_byte_idx = unwind_byte_idx;
+        return GRPC_ENDPOINT_PENDING;
       } else {
         /* TODO(klempner): Log some of these */
-        slice_state_destroy(state);
-        return GRPC_ENDPOINT_WRITE_ERROR;
+        return GRPC_ENDPOINT_ERROR;
       }
     }
 
-    /* TODO(klempner): Probably better to batch this after we finish flushing */
-    slice_state_remove_prefix(state, sent_length);
+    GPR_ASSERT(tcp->outgoing_byte_idx == 0);
+    trailing = sending_length - sent_length;
+    while (trailing > 0) {
+      ssize_t slice_length;
+
+      tcp->outgoing_slice_idx--;
+      slice_length = GPR_SLICE_LENGTH(
+          tcp->outgoing_buffer->slices[tcp->outgoing_slice_idx]);
+      if (slice_length > trailing) {
+        tcp->outgoing_byte_idx = slice_length - trailing;
+        break;
+      } else {
+        trailing -= slice_length;
+      }
+    }
 
-    if (!slice_state_has_available(state)) {
-      return GRPC_ENDPOINT_WRITE_DONE;
+    if (tcp->outgoing_slice_idx == tcp->outgoing_buffer->count) {
+      return GRPC_ENDPOINT_DONE;
     }
   };
 }
 
-static void grpc_tcp_handle_write(void *arg /* grpc_tcp */, int success) {
+static void tcp_handle_write(void *arg /* grpc_tcp */, int success) {
   grpc_tcp *tcp = (grpc_tcp *)arg;
-  grpc_endpoint_write_status write_status;
-  grpc_endpoint_cb_status cb_status;
-  grpc_endpoint_write_cb cb;
+  grpc_endpoint_op_status status;
+  grpc_iomgr_closure *cb;
 
   if (!success) {
-    slice_state_destroy(&tcp->write_state);
     cb = tcp->write_cb;
     tcp->write_cb = NULL;
-    cb(tcp->write_user_data, GRPC_ENDPOINT_CB_SHUTDOWN);
-    grpc_tcp_unref(tcp);
+    cb->cb(cb->cb_arg, 0);
+    TCP_UNREF(tcp, "write");
     return;
   }
 
   GRPC_TIMER_BEGIN(GRPC_PTAG_TCP_CB_WRITE, 0);
-  write_status = grpc_tcp_flush(tcp);
-  if (write_status == GRPC_ENDPOINT_WRITE_PENDING) {
+  status = tcp_flush(tcp);
+  if (status == GRPC_ENDPOINT_PENDING) {
     grpc_fd_notify_on_write(tcp->em_fd, &tcp->write_closure);
   } else {
-    slice_state_destroy(&tcp->write_state);
-    if (write_status == GRPC_ENDPOINT_WRITE_DONE) {
-      cb_status = GRPC_ENDPOINT_CB_OK;
-    } else {
-      cb_status = GRPC_ENDPOINT_CB_ERROR;
-    }
     cb = tcp->write_cb;
     tcp->write_cb = NULL;
-    cb(tcp->write_user_data, cb_status);
-    grpc_tcp_unref(tcp);
+    cb->cb(cb->cb_arg, status == GRPC_ENDPOINT_DONE);
+    TCP_UNREF(tcp, "write");
   }
   GRPC_TIMER_END(GRPC_PTAG_TCP_CB_WRITE, 0);
 }
 
-static grpc_endpoint_write_status grpc_tcp_write(grpc_endpoint *ep,
-                                                 gpr_slice *slices,
-                                                 size_t nslices,
-                                                 grpc_endpoint_write_cb cb,
-                                                 void *user_data) {
+static grpc_endpoint_op_status tcp_write(grpc_endpoint *ep,
+                                         gpr_slice_buffer *buf,
+                                         grpc_iomgr_closure *cb) {
   grpc_tcp *tcp = (grpc_tcp *)ep;
-  grpc_endpoint_write_status status;
+  grpc_endpoint_op_status status;
 
   if (grpc_tcp_trace) {
     size_t i;
 
-    for (i = 0; i < nslices; i++) {
-      char *data = gpr_dump_slice(slices[i], GPR_DUMP_HEX | GPR_DUMP_ASCII);
+    for (i = 0; i < buf->count; i++) {
+      char *data =
+          gpr_dump_slice(buf->slices[i], GPR_DUMP_HEX | GPR_DUMP_ASCII);
       gpr_log(GPR_DEBUG, "WRITE %p: %s", tcp, data);
       gpr_free(data);
     }
@@ -551,15 +387,19 @@ static grpc_endpoint_write_status grpc_tcp_write(grpc_endpoint *ep,
 
   GRPC_TIMER_BEGIN(GRPC_PTAG_TCP_WRITE, 0);
   GPR_ASSERT(tcp->write_cb == NULL);
-  slice_state_init(&tcp->write_state, slices, nslices, nslices);
 
-  status = grpc_tcp_flush(tcp);
-  if (status == GRPC_ENDPOINT_WRITE_PENDING) {
-    /* TODO(klempner): Consider inlining rather than malloc for small nslices */
-    slice_state_realloc(&tcp->write_state, nslices);
-    gpr_ref(&tcp->refcount);
+  if (buf->length == 0) {
+    GRPC_TIMER_END(GRPC_PTAG_TCP_WRITE, 0);
+    return GRPC_ENDPOINT_DONE;
+  }
+  tcp->outgoing_buffer = buf;
+  tcp->outgoing_slice_idx = 0;
+  tcp->outgoing_byte_idx = 0;
+
+  status = tcp_flush(tcp);
+  if (status == GRPC_ENDPOINT_PENDING) {
+    TCP_REF(tcp, "write");
     tcp->write_cb = cb;
-    tcp->write_user_data = user_data;
     grpc_fd_notify_on_write(tcp->em_fd, &tcp->write_closure);
   }
 
@@ -567,27 +407,25 @@ static grpc_endpoint_write_status grpc_tcp_write(grpc_endpoint *ep,
   return status;
 }
 
-static void grpc_tcp_add_to_pollset(grpc_endpoint *ep, grpc_pollset *pollset) {
+static void tcp_add_to_pollset(grpc_endpoint *ep, grpc_pollset *pollset) {
   grpc_tcp *tcp = (grpc_tcp *)ep;
   grpc_pollset_add_fd(pollset, tcp->em_fd);
 }
 
-static void grpc_tcp_add_to_pollset_set(grpc_endpoint *ep,
-                                        grpc_pollset_set *pollset_set) {
+static void tcp_add_to_pollset_set(grpc_endpoint *ep,
+                                   grpc_pollset_set *pollset_set) {
   grpc_tcp *tcp = (grpc_tcp *)ep;
   grpc_pollset_set_add_fd(pollset_set, tcp->em_fd);
 }
 
-static char *grpc_tcp_get_peer(grpc_endpoint *ep) {
+static char *tcp_get_peer(grpc_endpoint *ep) {
   grpc_tcp *tcp = (grpc_tcp *)ep;
   return gpr_strdup(tcp->peer_string);
 }
 
 static const grpc_endpoint_vtable vtable = {
-    grpc_tcp_notify_on_read, grpc_tcp_write,
-    grpc_tcp_add_to_pollset, grpc_tcp_add_to_pollset_set,
-    grpc_tcp_shutdown,       grpc_tcp_destroy,
-    grpc_tcp_get_peer};
+    tcp_read,     tcp_write,   tcp_add_to_pollset, tcp_add_to_pollset_set,
+    tcp_shutdown, tcp_destroy, tcp_get_peer};
 
 grpc_endpoint *grpc_tcp_create(grpc_fd *em_fd, size_t slice_size,
                                const char *peer_string) {
@@ -597,21 +435,19 @@ grpc_endpoint *grpc_tcp_create(grpc_fd *em_fd, size_t slice_size,
   tcp->fd = em_fd->fd;
   tcp->read_cb = NULL;
   tcp->write_cb = NULL;
-  tcp->read_user_data = NULL;
-  tcp->write_user_data = NULL;
+  tcp->incoming_buffer = NULL;
   tcp->slice_size = slice_size;
   tcp->iov_size = 1;
   tcp->finished_edge = 1;
-  slice_state_init(&tcp->write_state, NULL, 0, 0);
   /* paired with unref in grpc_tcp_destroy */
   gpr_ref_init(&tcp->refcount, 1);
   tcp->em_fd = em_fd;
-  tcp->read_closure.cb = grpc_tcp_handle_read;
+  tcp->read_closure.cb = tcp_handle_read;
   tcp->read_closure.cb_arg = tcp;
-  tcp->write_closure.cb = grpc_tcp_handle_write;
+  tcp->write_closure.cb = tcp_handle_write;
   tcp->write_closure.cb_arg = tcp;
 
-  tcp->handle_read_closure.cb = grpc_tcp_handle_read;
+  tcp->handle_read_closure.cb = tcp_handle_read;
   return &tcp->base;
 }